You are on page 1of 11

FEBRUARY 2003 THOMSON AND FINE 319

Estimating Mixed Layer Depth from Oceanic Profile Data


RICHARD E. THOMSON
Institute of Ocean Sciences, Sidney, British Columbia, Canada

ISAAC V. FINE
Institute of Ocean Sciences, Sidney, British Columbia, Canada, and Heat and Mass Transfer Institute,
National Academy of Sciences, Minsk, Belarus

(Manuscript received 29 December 2001, in final form 7 August 2002)

ABSTRACT
Estimates of mixed layer depth are important to a wide variety of oceanic investigations including upper-
ocean productivity, airsea exchange processes, and long-term climate change. In the absence of direct turbulent
dissipation measurements, mixed layer depth is commonly derived from oceanic profile data using threshold,
integral, least squares regression, or other proxy variables. The different methodologies often yield different
values for mixed layer depth. In this paper, a new methodthe split-and-merge (SM) methodis introduced
for determining the depth of the surface mixed layer and associated upper-ocean structure from digital conduc-
tivitytemperaturedepth (CTD) profiles. Two decades of CTD observations for the continental margin of British
Columbia are used to validate the SM method and to examine differences in mixed layer depth estimates for
the various computational techniques. On a profile-by-profile basis, close agreement is found between the SM
and de facto standard threshold methods. However, depth estimates from these two methods can differ significantly
from those obtained using the integral and least squares regression methods. The SM and threshold methods
are found to approximate the true mixed layer depth whereas the integral and regression methods typically
compute the depth of the underlying pycnocline. On a statistical basis, the marginally smaller standard errors
for spatially averaged mixed layer depths for the SM method suggest a slight improvement in depth determination
over threshold methods. This improvement, combined with the added ability of the SM method to delineate
simultaneously ancillary features of the upper ocean (such as the depth and gradient of the permanent pycnocline),
make it a valuable computational tool for characterizing the structure of the upper ocean.

1. Introduction Muller 1990) and by 100 m over an annual cycle (Do-


dimead et al. 1963; Sprintall and Roemmich 1999).
Turbulence generated in the ocean by the wind, con- Numerous studies have emphasized the importance
vective cooling, breaking waves, current shear, and other of mixed layer depth (MLD) estimation to upper-ocean
physical processes creates a surface layer characterized processes and variability (cf. Curry and Roy 1989; Rob-
by uniform to near-uniform density, active vertical mix- inson et al. 1993; Wijesekera and Gregg 1996; Kara et
ing, and high dissipation (e.g., Wijesekera and Gregg al. 2000a,b). Much of our physical understanding of
1996). The depth of this mixing layer (Brainerd and MLDs stems from an interest in transient, short-term
Gregg 1995) is ultimately determined by a balance be- (,1 day) fluctuations in sea surface temperature and the
tween the destabilizing effects of mechanical mixing accompanying exchanges of heat, carbon dioxide, and
and the stabilizing effects of surface buoyancy flux. Be- other properties with the overlying atmosphere [cf. the
cause of temporal variations in turbulence intensity and review by Large et al. (1994) and numerical modeling
buoyancy flux, the mixing layer may be embedded in work by Garwood (1977), Martin (1985), and Kantha
a deeper mixed layer of almost identical density to and Clayson (1994)]. In these studies, the mixed layer
the mixing layer and representing the time-integrated referred to as the diurnal mixed layer (DML) by Smyth
response to previous mixing events. Mixed layer depth et al. (1996a,b)is defined as the uppermost layer of
can vary by tens of meters over a diurnal cycle (Price uniform density, high turbulent dissipation, and diurnal
et al. 1986; Denman and Gargett 1988; Schneider and restratification. The commonly used threshold differ-
ence method (cf. Table 1) defines the base of the mixed
layer as the depth at which the potential density, su , in
Corresponding author address: Dr. Richard E. Thomson, Institute
of Ocean Sciences, P.O. Box 6000, 9860 West Saanich Rd., Sidney, the upper layer changes by 0.01 kg m 23 relative to the
BC V8L 4B2, Canada. ocean surface density. Although this definition is prob-
E-mail: thomsonr@pac.dfo-mpo.gc.ca ably too narrow for biological applicationsit often

q 2003 American Meteorological Society


320 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 20

TABLE 1. Selected definitions for mixed layer depth and other upper-ocean features. Here, T is temperature, S is salinity, and su is
potential density. The trapping depth, D T (Price et al. 1986), is analogous to the integral depth scale [Eq. (2.4)].

Source Name Definition

Price et al. (1986) Trapping depth, DT DT 5 DT 21 Ezi


z

T dz

Peters et al. (1989) Mixed layer depth, MLD Dsu 5 0.01 kg m23 (relative to z 5 0 m)
Schneider and Muller (1990) Mixed layer depth, MLD Dsu 5 0.01 kg m23 (relative to z 5 2.5 m);
Dsu 5 0.03 kg m23 (relative to z 5 2.5 m)
Wijffels et al. (1994) Mixed layer depth, MLD; Dsu 5 0.01 kg m23 (relative to z 5 2.5 m);
top of thermocline, TTC ]su /]z 5 0.01 kg m24
Brainerd and Gregg (1995) Mixed layer depth, MLD Dsu 5 0.0050.5 kg m23;
]su /]z 5 0.00050.05 kg m24
Smyth et al. (1996a,b) Diurnal mixed layer, DML; Dsu 5 0.01 kg m23;
upper-ocean layer, UOL su , 22 kg m23 (top of pycnocline)
Weller and Plueddemann (1996) Mixed layer depth, MLD; DT 5 0.018C (relative to z 5 2.25 m);
isopycnal layer depth, ILD; Dsu 5 0.03 kg m23 (relative to z 5 10 m);
seasonal thermocline depth, STD Dsu 5 0.15 kg m23 (relative to z 5 10 m)
Wijesekera and Gregg (1996) MLD Dsu 5 0.01 kg m23 (relative to z 5 0 m);
MLD1 ]su /]z 5 0.01 kg m24;
MLD2 ]su /]z 5 0.025 kg m24;
MLD3 ]S/]z 5 0.01 psu m21
Skyllingstad et al. (1999) MLD Dsu 5 0.01 kg m23 (relative to z 5 0 m)

neglects the underlying water of near-identical density less reliable because overturning turbulence can result
encompassing high chlorophyll, nutrient, and particle in adiabatic changes of 0.04 kg m 23 over depths of 10
concentrations (e.g., Robinson et al. 1993; Washburn et m (Schneider and Muller 1990).
al. 1998) and ignores the fact that the retreat of tur- This paper has a twofold purpose: 1) to introduce a
bulent mixing to shallower depths proceeds faster than split-and-merge (SM) algorithm for estimating the
the erosion of the stratification at the base . . . (Kara depths of the surface mixed layer and underlying fea-
et al. 2000a)the threshold difference method with the tures from oceanic profile data and 2) to compare results
specific threshold of 0.01 kg m 23 has emerged as the for the different estimators of mixed layer depth pres-
de facto standard by which different methods and stud- ently in use based on several decades of profile data
ies are compared. collected off the west coast of Canada. Formulation of
Methods for estimating the MLD from profile data the SM method originated from our need for a realistic
fall into three broad categories (cf. Schneider and Muller annual cycle in mixed layer depth for ecosystem models
1990; also, Table 1): 1) threshold methods, which find being developed for the west coast of Vancouver Island,
a predefined step in the surface profile (Price et al. 1986; British Columbia (Robinson et al. 1993; Robinson and
Lukas and Lindstrom 1991; Peters et al. 1988) or a
Ware 1994). A report on these models (C. Robinson
critical gradient for the upper layer (Lukas and Lind-
1997, personal communication) concluded that, In
strom 1991); 2) least squares regression methods, which
fit two or more line segments to near-surface profiles general, the upwelling rate and the minimum depth of
(Papadakis 1981, 1985); and 3) integral methods, which the summer mixed layer depth (off southwest Vancouver
calculate a depth scale for the upper layer (Freeland et Island) result in the greatest variability in annual plank-
al. 1997). Due to a lack of salinity data, or because ton or fish production. For our purpose, we found ex-
salinity data tend to be noisier than temperature data, isting techniques for estimating mixed layer depth either
the MLD is sometimes linked to a steplike change in too tightly constrained (e.g., the standard threshold
water temperature, with specified steps in the range method) or too loosely defined (e.g., the integral depth-
0.0180.58C (e.g., Levitus 1982; Weller and Pluedde- scale method). The SM method is shown to be a com-
mann 1996; Kara et al. 2000a,b). Where possible, it is putationally efficient and flexible curve-fitting technique
preferable to use potential density or sigma t (s t ) as that combines accurate estimation of mixed layer depth
estimators for MLD, primarily because it is the density with the versatility to define other upper-ocean features,
structure that directly affects the stability and degree of such as the depth of the top of the thermocline and the
turbulent mixing in the water column. In situ density is gradient of the main pycnocline.
FEBRUARY 2003 THOMSON AND FINE 321

2. Conventional methods fit and the Newtonian approximation method to find


a minimum variance solution to the general mixed
A common approach for estimating the surface mixed layer depth problem; Freeland et al. (1997) used a
layer depth is to specify a threshold slope for the vertical two-segment least squares approach to obtain a time
density gradient (e.g., Price et al. 1986; Lukas and Lind- series of winter mixed layer depth at Ocean Weather
strom 1991) or a threshold for the density offset relative Station P (508N, 1458W) in the northeast Pacific. The
to the ocean surface (e.g., Peters et al. 1989; Schneider two-segment case (Fig. 1c) can be solved analytically
and Muller 1990; Wijffels et al. 1994; Weller and Plued- and is useful for pre-CTD data. The three-segment
demann 1996). The density gradient method is consid- case requires special techniques (e.g., Papadakis
ered a less consistent estimator of MLD than the density 1985) and solutions can be unstable.
difference approach (Schneider and Muller 1990). Be- The two-segment approach seeks a least squares ap-
cause of the eddylike scales of turbulent mechanical proximation to a steplike water density, f (z), such that
mixing, density profiles are often piecewise-linear rather

5ss
than smoothing varying. As a consequence, MLD also u1 za , z , D
can be equated to the uppermost line segment in least f (z) 5 (2.1)
u2 D , z , zb ,
squares approximations to steplike density profiles (Pa-
padakis 1981; Freeland et al. 1997) or to the integral where z a is a near-surface depth; D is the estimated
depth scale for the surface layer (Price et al. 1986; Free- mixed layer depth; z b 5 200500 m is an arbitrary
land et al. 1997). Papadakis (1981, 1985) used a three- depth below the depth of seasonal mixing; and su1 ,
line approximation to estimate the MLD in terms of su2 are constant potential densities for the mixed layer
abrupt changes (break points) in the profiles. and intermediate layer, respectively. Minimizing the
integral
a. Threshold method
In airsea interaction studies (e.g., Wijesekera and
F5 E D

za
[su (z) 2 su1 ] 2 dz

E
Gregg 1996; Smyth et al. 1996a,b), the depth of the zb
surface mixed layer is defined as that depth z at which
1 [su (z) 2 su 2 ] 2 dz (2.2)
the potential density difference D su (z) 5 su (z) 2 su (z 0 ) D
in the upper ocean exceeds a specified threshold value,
commonly 0.01 kg m 23 (Fig. 1a); here z 0 is a reference leads to three equations for su1 , su2 , and D with the
depth (generally in the range z 0 5 0, the ocean surface, solution

E E
to 10 m), and su (z) 5 r u (z) 2 1000 kg m 23 is the density D zb
anomaly for measured potential density, r u . [For the su (z) dz su (z) dz
less-used threshold gradient method, mixed layer depth 1
F(D) 5 su (D) 2 5 0,
za D
is the depth at which the density gradient ]su /]z first 1
2 D 2 za zb 2 D
exceeds 0.01 kg m 24 (cf. Fig. 1b) or other specified
level.] Because the threshold method is comparatively (2.3)
simpledepth estimates can be made by hand without which can be solved numerically. Two-segment ap-
analytical computationsand because the threshold dif- proximations are computationally stable. Increasing the
ference of 0.01 kg m 23 generally yields diurnal depth number of segments or steps can improve the ap-
estimates similar to those from turbulent dissipation proximation to the profile data but typically leads to
measurements, the threshold difference method with the greater complexity. The quality of the least squares fit
threshold 0.01 kg m 23 has become the de facto standard varies from profile to profile depending on the structure
for many mixed layer depth studies. However, in ocean- of the underlying layering.
climate studies, where the focus is on the variability in
MLD averaged over periods of months and longer,
threshold values in excess of 0.125 kg m 23 are more c. Integral depth-scale method
commonly used for monthly mean data (cf. Table 1 in
A simple estimate of the mixed layer depth is the
Kara et al. 2000a).
integral depth scale (Fig. 1d) [also called the trapping
depth by Price et al. (1986)]:

E E
b. Step-function least squares regression method zb zb

Problems with profile approximations have led to zN 2b (z) dz (su b 2 su ) dz

E
0 za
the formulation of customized curve-fitting algo- D5 5 , (2.4)
rithms for oceanic profiles, including the use of least
zb
su b 2 su a
squares linear approximations (Papadakis 1981; Free- N (z) dz
2
b
0
land et al. 1997) and form oscillators (Papadakis
1985). Papadakis (1981) used a three-segment linear where, as in the previous section, z a and z b are a near-
322 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 20

FIG. 1. Estimation of mixed layer depth (D) for different methods for the density profile obtained for station LH07 on the west coast of
Vancouver Island on 19 Jul 1997 (see Fig. 3 for station location). (a) Threshold difference method with standard density step D su (z) 5 su (z)
2 su (0) 5 0.01 kg m 23 ; (b) threshold gradient method for ]su /] z 5 0.01 kg m 24 ; (c) two-segment least squares method [cf. Eqs. (2.1)
(2.3)] for z a 5 10 m, z b 5 200 m, constant su1 and su2 , and s (D) 5 (su1 1 su2 )/2; and (d) integral depth scale, D, with z a 5 10 m and z b
5 200 m [Eq. (2.4)]. In (c), each of the paired shaded regions have equal positive (1) and negative (2) areas. In (d), the darkly shaded
rectangle has sides of length D and L s 5 su (z b ) 2 su (z a ), with the area D 3 L of the rectangle equal to the integral I 5 # zb [su (z) 2 sub ]
z

dz denoted by the shaded region to the right of the density profile.

surface depth and an arbitrary reference depth, and sub upper portion of a density profile, the step-function and
5 su (z b ), sua 5 su (z a ) (cf. Freeland et al. 1997). Here, integral-depth approaches require specification of a deep
reference density that is generally much deeper than the

1 2
g dru
1/2
mixed layer depth.
N b (z) 5 2 (2.5)
r 0 dz
is the buoyancy (BruntVaisala) frequency, g is the ac- 3. The split-and-merge algorithm
celeration of gravity, and r 0 is a reference density. Un- Application of the threshold difference method to the
like the threshold methods, which often require only the 4811 CTD profiles collected from Canadian Department
FEBRUARY 2003 THOMSON AND FINE 323

or steplike offsetsare adjusted to fit the available data


(Fig. 2). The algorithm provides profile decomposition
by defining the locations of the break points separating
the different segments, together with the piecewise ap-
proximation parameters and the error of the approxi-
mation. In general, the fitted segments are disjointed but
can be made to satisfy a continuity requirement by a
retrospective local adjustment of the approximating
curves. We consider the problem of fitting a series of
segments to the profile, f (z), where z is depth and f
represents temperature, salinity, density, fluorescence,
or other profile variable. Given a set of points S 5 {z i ,
f i }, i 5 1, 2, . . . , N, we seek the minimum number n
such that S is divided into n subsets S1 , S 2 , . . . , S n in
which the data points for each subset are approximated
by a polynomial of order at most m 2 1 with an error
norm that is less than some specified quantity . Here,
error norm is the root-mean-square (rms) difference,
maximum difference, or other statistical measure of the
difference between the observed and fitted curves for a
given subset (cf. the appendix). The SM algorithm
merges adjacent fitted segments with similar approxi-
mating coefficients and splits those segments with un-
acceptable error norms. A least squares method, or sim-
ilar approach, is used to fit the curve to the specified
FIG. 2. Illustration of the SM algorithm applied to a case where dataset. Because fitting higher-order polynomials has its
the optimum segmentation can be found in single iterations (after own set of problems, we restrict our discussion to piece-
Pavlidis and Horowitz 1974). The initial fit (dashed line) to the left- wise-linear approximations for which m 5 2. The up-
hand segment of the curve f (z) is split at the break point to form
three separate segments. The two right-hand segments are then permost segment (representing the mixed layer) is as-
merged to form one segment. sumed to have a near-uniform vertical distribution so
that m 5 1 for this segment.
The SM method removes the need for the careful a
of Fisheries and Oceans research vessels working off priori choice of the number of segments n as well as
the west coast of Vancouver Island from 1978 to 2000 the need to specify the initial segmentation. In this way,
yielded instances where the estimated mixed layer depth it is straightforward to obtain segmentation where the
differed markedly from the visually estimated mixed error norm on each segment (or over all segments) does
layer depth (i.e., the depth to the top of the first pyc- not exceed a specified bound. Computational time is of
nocline estimated from density plots). This, and our con- order N. For a specified error norm, we equate the mixed
clusion that depth estimates from the step-function and layer depth D with the uppermost (constant) segment
integral-scale methods were more representative of the of the piecewise-linear fit. (See the appendix for further
main pycnocline depth rather than the surface mixed details on the SM method.)
layer depth, caused us to seek a more reliable method To avoid scaling problems for profile variables, in-
for estimating mixed layer depth. Moreover, because the cluding the need for a different error norm for each
separation of profile data into distinct layers is of ocean- variable, we have normalized both z and f (z) such that
ographic importance, we wanted a method that could
simultaneously approximate other upper-ocean struc- z 2 z min f (z) 2 fmin
z* 5 ; f*(z) 5 , (3.1)
tural features in addition to the mixed layer depth. For z max 2 z min fmax 2 fmin
these applications, we turned to the SM algorithm de-
veloped by Pavlidis and Horowitz (1974) to estimate where subscripts denote the maximum and minimum
the optimal decomposition of plane curves and wave- values of the given variable over the depth range of
forms. interest and 0 # [z*, f*(z)] # 1. Following other au-
thors (see Table 1), we have used a nonzero starting
depth (here, zmin 5 2.5 m), to avoid the effects of prop-
a. Methodology
wash and turbulent flow past the hull of the ship during
The SM algorithm approximates a specified curve station keeping, and zmax 5 150 m to ensure capture of
using piecewise polynomial functions (or other specified the mixed layer depth regardless of season. As with the
functions) in which the break points in the fitted curve threshold method, the SM algorithm requires a prede-
locations of subset boundaries such as changes in slope fined error norm . To determine the sensitivity of MLD
324 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 20

estimates to the specified error norm, we examined es- Table 2a presents a statistical summary of the mean
timates over a wide range of values, from 0.001 to depths and standard deviations for all five methods for
0.03, and found that results for 5 0.01 best corre- all density profiles collected off the west coast of Van-
sponded to our visual MLD estimates. This error couver Island. Results for the threshold difference meth-
norm also coincides closely with accepted threshold val- od (TH) are based on the threshold D su 5 0.01 kg m 23
ues for the established threshold method, giving further and those for the threshold gradient method (GR) on a
support for the SM approach. We note that 5 0.01 gradient D su /Dz 5 0.006 kg m 24 . The specified error
yields a mean MLD of 15.4 m, the same mean MLD norm 5 0.003 for the SM method coincides with the
obtained for the threshold D su 5 0.03 kg m 23 . The threshold difference of 0.01 kg m 23 . As the tabulated
smaller error norm 5 0.003 yields a slightly smaller results indicate, the mean mixed layer depths obtained
mean MLD of 12.4 m, which is similar to the mean using the threshold and SM methods are markedly dif-
MLD obtained from the standard threshold D su 5 ferent from those based on the regression and integral
0.01 kg m 23 . We therefore use the error norm 5 0.003 methods. Mean depths from the threshold and SM meth-
when making direct comparisons of the SM and standard ods (MLD 12 m) are a factor of 5 smaller than those
threshold methods. for the regression and integral methods (MLD 60 m),
Figure 3 provides examples of SM estimates of mixed and most accurately match visual estimates of the uni-
layer depth for CTD profiles collected in July 1997 form-density surface mixed layer. Estimates from the
along ship survey lines B, D, H, and M off the west regression and integral methods are more representative
coast of Vancouver Island (see insert in Fig. 3). In almost of the permanent pycnocline depth than the surface
all cases, the depth estimates for the SM method with mixed layer depth.
5 0.01 are consistent with our visual estimations of Because the mean MLD estimates determined by the
the mixed layer depth. Similar findings are obtained for threshold and SM methods are nearly identical, we use
different lines and different summer survey data. a second-order statistic (the standard error) as a measure
of the relative performance or predictive skill of a
given method. The relatively small standard error
b. Comparison of methods
(5standard deviation divided by (N 2 1), where N
All CTD profiles collected seaward of Vancouver Is- is the number of estimates) in mixed layer depth for the
land from 1978 to 2000 have been used to compare threshold and SM methods (Table 2a) confirms the abil-
results from the SM method with those from conven- ity of these methods to closely track short-period chang-
tional methods. Figure 4a presents an example of a well- es in layer depth. In contrast, the regression and integral
defined mixed layer in which the upper layer has uni- methodswhich provide a weighted depth for the main
form density and is underlaid by a sharp pycnocline. pycnoclineappear to be more affected by large-am-
For this example, the threshold and SM methods give plitude internal tides whose vertical displacements are
nearly identical MLD results over a wide range of typically largest within the pycnocline. The three-layer
threshold and error norm values. We note, however, that (D 3 ) regression method performed better than the two-
for the threshold method the MLD estimate increases layer (D 2 ) regression method, but it too proved unsat-
with increased threshold while for the SM method, the isfactory when the depth profile below the surface mixed
MLD estimate remains unchanged over a wide range of layer became structurally complicated. According to Ta-
norm values (i.e., is less sensitive to the choice of error). ble 2b (which also includes a three-step SM analysis
The underlying pycnocline is closely approximated by using n 5 3), the covariance between the two-step re-
the SM method, demonstrating the methods consider- gression and integral methods is relatively high (r 2 5
able flexibility. 0.71), whereas the covariance between these methods
Figure 4b presents a more complicated case in which and other methods is low (r 2 , 0.15). For the threshold
the upper layer has a weak density gradient. Here, the and SM methods, within-group correlation is highest
ability of the threshold method to estimate the MLD is between the threshold difference and SM methods (r 2
strongly dependent on the threshold value. In contrast, 5 0.90) and lowest between the SM and threshold gra-
the SM algorithm finds an MLD that corresponds closely dient methods (r 2 5 0.76). Thus, based on computed
to the visual estimate. When the upper layer has a sig- means, standard errors, and covariances, we conclude
nificant density gradient (as in Fig. 4c), both the that only the threshold and SM methods provide con-
threshold and SM methods have difficulty estimating a sistently accurate estimates of mixed layer depth.
mixed layer depth. In such cases, there is no well-de- Because the error norm 5 0.003 for the SM method
fined mixed layer and depth estimates invariably depend was selected to provide MLD estimates that were com-
on the specified error values. In particular, if the profile patible with depth estimates obtained using the standard
has a measurable pycnocline in the upper portion of the threshold and threshold gradient methods, the mean
profile, the SM algorithm will find the start depth of the depths for all methods (12.4 m for the SM method and
layer, which in our case has been set to 2.5 m (Fig. 4c). 12.5 m for the threshold difference and threshold gra-
Similar results were found scattered throughout the 21- dient methods) are essentially identical. However, the
yr dataset. variance in the mixed layer depth estimateswhich
FEBRUARY 2003 THOMSON AND FINE 325

FIG. 3. Top portion of 1-m average density (su ) profiles derived from CTD casts collected in Jul 1997 off the west coast of Vancouver
Island (see map). Horizontal bars denote mixed layer depth derived using the SM algorithm. Profiles of su typically range from 22 to 25
kg m 23 . A scale of three su units is shown in the line B profile plots.
326 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 20

FIG. 4. The 1-m average density profile data (solid dots) and line segments fitted using the SM method for line B for 24 Jul 1997 (for
location, see Fig. 3). Depth estimates for the SM method use 5 0.01 and those for the threshold difference method (TH) use both D su 5
0.01 and 0.03 kg m 23 . (a) A well-defined mixed layer, and underlying pycnocline; (b) less well-defined mixed layer and pycnocline; and
(c) a poorly defined (essentially, nonexistent) mixed layer and pycnocline. Time and positional coordinates for each profile are shown. To
avoid prop-wash and ship-wake contamination, the top 2.5 m of each profile has been omitted.

originates with internal wave motions, measurement 4. Discussion and summary


limitations, and other factorsdiffers among the three
methods. Our contention that the SM method outper- The split-and-merge algorithm provides an accurate
forms the two threshold methods is based primarily and flexible new method for estimating the depth of the
on the smaller standard error for the SM method (Table surface mixed layer from oceanic profiles. The method
2a) and on the smaller percentage change in mixed layer eliminates the need for good initialization of the com-
depth caused by the incremental changes in the applied putation and allows for the fit of a variable number
error norm relative to corresponding changes in the of linear segments to the profile data. Along with es-
threshold levels (Fig. 5). Twenty-five estimates were timates of mixed layer depth, the method can be used
used to construct each of the curves in Fig. 5, with to delineate other structural features of the upper ocean
values increasing from 0.004 kg m 23 in steps of 20% such as the gradient of the main pycnocline and the
for the threshold difference method, from 0.0015 kg m 24 depth to the top of the thermocline (TTC).
in steps of 15% for the threshold gradient method, and Our intercomparison among the various depth esti-
from 0.001 in steps of 20% for the SM method. As the mators (the SM method, two threshold-type methods,
thresholds and error norm are increased, estimates of the two- and three-step regression methods, and the in-
the mean mixed layer depth increase (from roughly 10 tegral-scale method) demonstrates that, in the absence
to 20 m) and the standard error decreases. Assuming of turbulent dissipation measurements, only the thresh-
that true mixed layer depth variability due to internal old and SM methods provide accurate estimates of
waves and other factors contributes equally to the mixed layer depth. This is clearly borne out by the mixed
threshold and SM methods, we conclude that the thresh- layer depth estimates obtained for two decades of den-
old methods have greater variance than the SM method sity profiles collected off the west coast of Vancouver
because of higher computational inaccuracy. For the er- Island. According to results in Table 2a, the ;60 m
ror norm ranges considered, the SM method is a slightly mean depth obtained from the two-step regression and
better estimator. integral-scale methods (which is roughly five times the

TABLE 2a. The mean mixed layer depth (MLDsub ) and standard error (std dev divided by N, N 5 number of samples) for different
estimation methods. Subscripts SM, TH, and GR denote the SM, threshold difference, and threshold gradient methods, respectively; Ds is
the integral depth scale, D 2 is the MLD from the two-step function, and D 3 for the three-step function approach. The SM method uses an
error norm e 5 0.003.

MLDSM MLDTH MLDGR Ds D2 D3


Mean (m) (12.51 6 0.32) (12.43 6 0.33) (12.57 6 0.38) (60.4 6 0.6) (60.9 6 0.7) (35.60 6 0.46)
FEBRUARY 2003 THOMSON AND FINE 327

TABLE 2b. The correlation-squared (r 2 ) matrix. (See Table 2a


for terminology.)

MLDSM MLDTH MLDGR Ds D2 D3


MLDSM 1 0.90 0.76 0.28 0.09 0.11
MLDTH 1 0.79 0.29 0.10 0.12
MLDGR 1 0.30 0.10 0.13
Ds 1 0.71 0.29
D2 1 0.13

;12 m mean depth from the SM and threshold methods,


and three times the ;35 m mean depth for the three-
step regression method) is actually an estimate of the
permanent pycnocline depth rather than the depth of the
much shallower surface mixed layer. Similar results are
obtained when the MLD data are analyzed according to
season (Fig. 6). For the SM and threshold methods, the
seasonal-mean MLD is shallower and associated con-
fidence intervals more tightly constrained with depth
for all seasons compared with the integral and regression
methods. The normal MLD confidence distributions for
the regression and integral methods are more reminis-
cent of random processes than natural variability as-
sociated with the spatial and temporal variations in FIG. 5. Relative error (5 std dev in depth estimate divided by
MLD. (N 2 1)) for N estimates calculated as the percentage of the mean
mixed layer depth (summer-only data, west coast of Vancouver Is-
Of the two most accurate methods, the SM method land): GR 5 threshold gradient method; TH 5 threshold difference
yields a marginally lower standard error than the thresh- method; and SM 5 split-and-merge method. Each curve covers a
old difference method. This error arises from compu- broad range of 25 error norm (or threshold) values. The vertical line
tational inaccuracy as well as from temporal and spatial denotes the mean mixed layer depth for the de facto standard threshold
variability in mixed layer depth arising from surface difference, D su 5 0.01 kg m 23 .
heating/cooling, internal wave displacements, and other
physical processes. If we assume that smaller standard 1) find the smallest number of subsets n such that the
error is synonymous with improved computational skill, specified error norm, the statistical measure used to de-
then the SM method slightly outperforms the thresh- fine the difference between the measured and fitted
old difference method in its ability to determine the curves, does not exceed a given bound, or 2) minimize
surface mixed layer depth (the SM and threshold meth- the error norm for a given n. Approach 1 usually in-
ods would yield the same variance in mixed layer depth volves a linear scan of the data points to be approxi-
if it were not for the slightly better computational ac- mated and has a computational time of order N logN or
curacy of the SM method). In effect, the SM method N 2 (Ramer 1972). Approach 2 begins with small incre-
provides an additional tool for estimating mixed layer ments in the locations of the break points (subset
depth from profile data. The greater computational com- boundaries where there are abrupt changes in slope)
plexity of the SM method is compensated for by its and, in the worst case, requires about N iterations. The
slightly greater reliability and its added ability to si- worst case occurs when all the break points are clustered
multaneously determine additional structural features of at the wrong end and have to be moved across the
the upper ocean. entire array of data points (Pavlidis and Horowitz 1974).
For most other methods, initialization is critical, but is
Acknowledgments. We thank Alexander Rabinovich, itself a nontrivial problem. The SM algorithm not only
Evgueni Kulikov, Ken Denman, and several anonymous eliminates the need for good initialization but also
reviewers for their valuable comments on the manu- allows for a variable number of segments. In effect, the
script, and Patricia Kimber for her assistance with the method attacks the problem of approach 1 using the
figures. algorithms for approach 2, with a significant decrease
in computing time.
APPENDIX The task is to determine the minimum number n such
that the set of points S is divided into n subsets S1 , S 2 ,
Steps in the SM Method
. . . , S n , where, for each subset, the data points are ap-
The problem of finding the best piecewise-linear proximated by a linear polynomial of the form f k (z) 5
approximation to the N-point curve S 5 {(z i , f i ), i 5 a k 1 b k z, k 5 1, 2, . . . , n with an error of less than
1, 2, . . . , N} generally proceeds in one of two ways: some prescribed quantity and where the coefficients
328 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 20

FIG. 6. Seasonal mean mixed layer depths (solid dots) and associated percentage confidence intervals based on CTD data for the west
coast of Vancouver Island for the period 19782000. Winter 5 JanFebMar, spring 5 AprMayJune, summer 5 JulyAugSep, and fall
5 OctNovDec. GR 5 threshold gradient method, TH 5 threshold difference method, SM 5 split-and-merge method, Int 5 integral
method, RG2 5 two-step regression, and RG3 5 three-step regression.

a k and b k are found using the least squares method. For segments with large error norms and to merge adjacent
the first subset (here, the mixed layer), the constant b1 segments with similar approximating coefficients (i.e.,
5 0. The choice of the error norm , which defines how split and merge). The fundamental steps in the spit-
well the fitted curves match the observations, depends and-merge method are as follows:
to some extent on the application and is commonly de-
fined in two ways: 1) through the integral square error Step 1 (the split procedure)For an arbitrary initial
k(1) where subset k 5 1, 2, . . . , n r , obtain the polynomial fit
k(1) 5 O ,2
ik for (z i , f i ) S k ,
and check to see if the error norm k for this fit
exceeds the imposed limit, . (Here, we can use n r
5 2 since the program automatically adjusts to find
i

and 2) by the maximum error k(2) where the appropriate number of subsets for the specified
error norm.) If k . , split the kth subset into two
k(2) 5 max( ik ), for (z i , f i ) S k
parts and increment the number of subsets to n r 1
and where ik 5 | f i 2 f k (z i ) | . Sensitivity to noise is 1. The dividing point of the kth subset is determined
an important difference between the k(1) and k(2) norms. according to the following rule: If there are two or
Methods used to find the best piecewise-linear poly- more points where the pointwise error is maximum,
nomial approximation to a curve almost always involve then use as a dividing point the midpoint between
a linear scan of the points to be approximated, possibly a pair of them. Otherwise, divide the segment S k
with repetitions. The process of break-point adjustment in half. Calculate the error norms k on each new
can be accelerated substantially if one is allowed to split interval.
FEBRUARY 2003 THOMSON AND FINE 329

Step 2 (the merge procedure)For k 5 1, 2, . . . , n r , Lukas, R., and E. Lindstrom, 1991: The mixed layer of the western
equatorial Pacific Ocean. J. Geophys. Res., 96, 33433358.
merge segments S k and S k11 provided that this will Martin, P. J., 1985: Simulation of the mixed layer at OWS November
result in a new segment with k # . Then, decrease and Papa with several models. J. Geophys. Res., 90, 903916.
n r by unity to n r 2 1 and calculate the error norm Papadakis, J. E., 1981: Determination of the wind mixed layer by an
on each new interval. extension of Newtons method. Pacific Marine Sci. Rep. 81-9,
Step 3 (the break-point adjustment procedure)For Institute of Ocean Sciences, Sidney, BC, Canada, 32 pp.
, 1985: On a class of form oscillators. Speculations Sci. Technol.,
k 5 1, 2, . . . , n r 2 1, for every adjacent pair of 8, 291303.
subsets (S k and S k11 ), scan the position of the break Pavlidis, T., and S. L. Horowitz, 1974: Segmentation of plan curves.
points to decrease the maximum of the error ( k , IEEE Trans. Comput., C-23, 860870.
k11 ). Peters, H., M. C. Gregg, and J. M. Toole, 1988: On the parameter-
ization of equatorial turbulence. J. Geophys. Res., 93, 1199
Step 4If no changes in the segments have occurred 1218.
in any of steps 13 above, then terminate. Or else , , and , 1989: Meridional variability of turbulence
go to step 1. through the equatorial undercurrent. J. Geophys. Res., 94,
18 00318 009.
The above procedure is an algorithm that terminates Price, J. F., R. A. Weller, and R. Pinkel, 1986: Diurnal cycling: Ob-
after a local minimum n, for the number of segments servations and models of the upper ocean response to diurnal
heating, cooling, and wind mixing. J. Geophys. Res., 91, 8411
is reached and the criterion for the termination is sat- 8427.
isfied. The algorithm computational time is linear in N, Ramer, U., 1972: An iterative procedure for the polygonal approxi-
even for the worst cases. mation of plane curves. J. Comput. Graphics Image Process.,
1, 244256.
Robinson, C. L. K., and D. M. Ware, 1994: Modeling pelagic fish
REFERENCES and plankton trophodynamics off southwestern Vancouver Is-
land, British Columbia. Can. J. Fish. Aquat. Sci., 51, 17371751.
, , and T. R. Parsons, 1993: Simulated annual plankton
Brainerd, K. E., and M. C. Gregg, 1995: Surface mixed and mixing production in the northeastern Pacific coastal upwelling domain.
layer depths. Deep-Sea Res., 42A, 15211543. J. Plankton Res., 15 (2), 161183.
Curry, P., and C. Roy, 1989: Optimal environmental window and Schneider, N., and P. Muller, 1990: The meridional and seasonal struc-
pelagic fish recruitment success in upwelling areas. Can. J. Fish. tures of the mixed-layer depth and its diurnal amplitude observed
Aquat. Sci., 46, 670680. during the Hawaii-to-Tahiti Shuttle experiment. J. Phys. Ocean-
Denman, K. L., and A. E. Gargett, 1988: Multiple thermoclines are ogr., 20, 13951404.
barriers to vertical exchange in the subarctic Pacific during SU- Skyllingstad, E. D., W. D. Smyth, J. N. Moum, and H. Wijesekera,
PER, May 1984. J. Mar. Res., 46, 77103. 1999: Upper-ocean turbulence during a westerly wind burst: A
Dodimead, A. J., F. Favorite, and T. Hirano, 1963: Salmon of the comparison of large-eddy simulation results and microstructure
North Pacific Ocean. Part II: Review of oceanography of the measurements. J. Phys. Oceanogr., 29, 528.
subarctic Pacific region. International North Pacific Fisheries Smyth, W. D., D. Hebert, and J. N. Moum, 1996a: Local ocean re-
Commission, Bull. 13, Vancouver, BC, Canada, 195 pp. sponse to a multiphase westerly windburst. Part 1: The dynamic
Freeland, H., K. Denman, C. S. Wong, F. Whitney, and R. Jacques, response. J. Geophys. Res., 101, 22 49522 512.
1997: Evidence of change in the winter mixed layer in the north- , , and , 1996b: Local ocean response to a multiphase
east Pacific Ocean. Deep-Sea Res., 44A, 21172129. westerly windburst. Part 2: Thermal and freshwater responses.
Garwood, R. W., 1977: An oceanic mixed layer model capable of J. Geophys. Res., 101, 22 51322 533.
simulating cyclic states. J. Phys. Oceanogr., 7, 455471. Sprintall, J., and D. Roemmich, 1999: Characterizing the structure
Kantha, L. H., and C. A. Clayson, 1994: An improved mixed layer of the surface layer in the Pacific Ocean. J. Geophys. Res., 104,
model for geophysical applications. J. Geophys. Res., 99, 23 29723 311.
25 23525 266. Washburn, L., B. M. Emery, B. H. Jones, and D. G. Ondercin, 1998:
Kara, A. B., P. A. Rochford, and H. E. Hurlburt, 2000a: An optimal Eddy stirring and phytoplankton patchiness in the subarctic
definition for ocean mixed layer depth. J. Geophys. Res., 105, North Atlantic in late summer. Deep-Sea Res., 45A, 14111439.
16 80316 821. Weller, R. A., and A. J. Plueddemann, 1996: Observations of the
, , and , 2000b: Mixed layer depth variability and bar- vertical structure of the oceanic boundary layer. J. Geophys. Res.,
rier layer formation over the North Pacific Ocean. J. Geophys. 101, 87898806.
Res., 105, 16 78316 801. Wijesekera, R. W., and M. C. Gregg, 1996: Surface layer response
Large, W. G., J. C. McWilliams, and S. C. Doney, 1994: Oceanic to weak winds, westerly bursts, and rain squalls in the western
vertical mixing: A review and a model with nonlocal boundary Pacific warm pool. J. Geophys. Res., 101, 977997.
layer parameterization. Rev. Geophys., 32, 363403. Wijffels, S., E. Firing, and H. Bryden, 1994: Direct observations of
Levitus, S., 1982: Climatological Atlas of the World Ocean. NOAA the Ekman balance at 108N in the Pacific. J. Phys. Oceanogr.,
Prof. Paper 13, 173 pp. and 17 microfiche. 24, 16661679.